[Data] Shared code changes for Iceberg [1/6]

[AarryaSaraf]

Description

First of a 6-PR stack porting read_iceberg to Datasource V2. This one
generalizes the shared DataSourceV2 abstractions so that a catalog-backed source
can be built on them — no Iceberg code lands here. All three changes are no-ops
for Parquet, the only DataSourceV2 in the tree today.

• DataSourceV2.schema_needs_file_sample lets a source whose schema comes
  from a catalog skip file sampling at plan time, rather than failing with
  no files found under ... on an empty listing. infer_schema and
  resolve_partitioning widen to Optional to accept the missing sample.

• ReadFiles.apply_predicate re-emits the residual from
  Scanner.push_filters as a Filter instead of dropping it, so a scanner that
  can translate only part of a predicate no longer silently returns rows the
  predicate rejects.

• PushdownCountFiles calls Scanner.metadata_row_count_is_exact() and
  FileIndexer.as_whole_file_indexer() instead of type-testing for
  ArrowFileScanner and NonSamplingFileIndexer. Both hooks default to "not
  eligible", so an unknown component declines the rewrite rather than risking a
  wrong count().

Related issues

N/A

Additional information

No new tests yet — the hooks get their coverage in [2/6], where the Iceberg
datasource is the first real implementation of all three.

[gemini-code-assist]

Code Review

This pull request refactors the PushdownCountFiles optimization rule to use abstract interfaces (metadata_row_count_is_exact and as_whole_file_indexer) instead of concrete type-testing, and introduces schema_needs_file_sample to allow schema inference without sampling files. Feedback highlights two key issues: first, failing to override as_whole_file_indexer in NonSamplingFileIndexer will cause a regression by silently disabling the count pushdown optimization; second, requiring self.partition_predicate is None in ArrowFileScanner is overly restrictive for partitioned datasets where partition filters do not reduce the row count of individual matched files.

[gemini-code-assist]

The default implementation of as_whole_file_indexer returns None. Since NonSamplingFileIndexer is not updated in this PR to override this method, it will inherit this default implementation and return None. This silently disables the PushdownCountFiles optimization for all existing file-based datasources (such as Parquet) that rely on NonSamplingFileIndexer.\n\nTo prevent this optimization regression, NonSamplingFileIndexer should override as_whole_file_indexer to return self.

[gemini-code-assist]

The check self.partition_predicate is None is unnecessarily restrictive. In Hive-partitioned datasets, partition columns are uniform across all rows within any given file. Partition pruning filters out entire files during listing, meaning that for any file that is actually read, 100% of its rows match the partition predicate. Therefore, the partition predicate does not reduce the row count within the selected files, and the metadata row count remains exact.\n\nRemoving self.partition_predicate is None allows the PushdownCountFiles optimization to run for partitioned queries (e.g., ds.filter(year=2023).count()), which otherwise would fall back to a full read of all Parquet files.

        return (
            self.predicate is None
            and self.limit is None
        )